Reproducibility of Rank-1 results after the OSS release: investigate a potential gap and/or update the README.

The current table in the README presents results for Rank-1 BNNs that date back to April/May which :
* used a different evaluation log-likelihood
* possibly had a slightly different CIFAR data loader
* mostly were 3 to 5-seed averages instead of our current standard of 10-seed averages

The latest results with Gaussian Rank-1 BNNs are slightly underperforming (by an order of 0.05%) on CIFAR-10 while outperforming on CIFAR-100 (by an order of 0.1-0.4%) and require investigating or simply updating the README.

Start with the README table and compare its Rank-1 BNN results with the latest Gaussian Rank-1 results. Check the evaluation log-likelihood, CIFAR data loader, and seed counts mentioned in the issue; done means the discrepancy is explained and the README is updated if needed.

Written by the indexing model from the issue text.